基于改进GoogLeNet的玉米叶片病害识别及其可解释性研究  

Identification and explainability of maize leaf diseases based on improved GoogLeNet

作　　者：牛潘婷 张宝林[1,2,3] 潘丽杰 郭建鹏 Niu Panting;Zhang Baolin;Pan Lijie;Guo Jianpeng(College of Chemistry and Environmental Sciences,Inner Mongolia Normal University,Hohhot,010020,China;Inner Mongolia Key Laboratory of Environmental Chemistry,Hohhot,010020,China;Inner Mongolia Water-saving Agriculture Engineering Research Center,Hohhot,010020,China)

机构地区：[1]内蒙古师范大学化学与环境科学学院,呼和浩特市010020 [2]内蒙古自治区环境化学重点实验室,呼和浩特市010020 [3]内蒙古节水农业工程研究中心,呼和浩特市010020

出　　处：《中国农机化学报》2025年第1期204-212,共9页Journal of Chinese Agricultural Mechanization

基　　金：内蒙古自然科学基金项目(2022LHMS03009);内蒙古自治区科技重大专项课题(2021ZD0003—1);内蒙古师范大学基本科研业务费专项资金(2022JBTD009)。

摘　　要：为加强农作物病害的识别,减少病害发生的频率与强度,提高农作物产量与品质,基于迁移学习构建5种深度学习网络,对玉米叶片锈病、大小斑病和灰斑病进行识别分类研究。通过对比AlexNet、VGG19、ResNet50、GoogLeNet和MobileNetV2深度学习网络,GoogLeNet的识别准确率最高,达到96.3%,模型收敛效果最好。通过进一步优化GoogLeNet模型架构,在inception模块中插入卷积注意力模块CBAM,使用LeakyReLU激活函数替换ReLU函数,改进后网络通道注意力增强,测试集的识别准确率达到99.0%,识别准确率提高2.7%。采用CAM和LIME算法对模型的可解释性分析,改进后网络的可解释性增强,更好地关注叶片病害部分。In order to strengthen the identification of crop diseases,reduce the frequency and intensity of disease occurrence,and increase crop yield and quality,five kinds of deep learning frameworks based on transfer learning was used to identify and classify maize leaf diseases,including corn rust,leaf blight and gray spot.Compared with AlexNet,VGG19,ResNet50,GoogLeNet and MobileNetV2 deep learning networks,the recognition accuracy of GoogLeNet is the highest by 96.3%,and the model convergence effect is the best.By further optimizing the GoogLeNet model architecture,the Convolutional Block Attention Module(CBAM)is inserted into the inception module,and the LeakyReLU activation function is used to replace the ReLU function.After the improvement,the network channel attention is strengthened,the recognition accuracy of the test dataset reaches 99.0%,and the recognition accuracy is increased by 2.7%.CAM(Class Activation Mapping)and LIME(Local Interpretable Model-Agnostic Explanations)algorithms are used to analyze the model interpretability,the interpretability of the improved network shows higher explainability,with more attention on disease affected leaf regions.

关 键 词：深度学习 玉米叶片病害 迁移学习 可解释性 图像处理 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S435.131[自动化与计算机技术—计算机科学与技术]